4.1 Physical basis of predictor selection for each EOF mode
Figure 2 shows distribution of the correlation coefficient between the PC1 and SST anomalies, geopotential height at 850hPa (hereafter H850) and VWS from both observational dataset and model hindcast during 1991–2020. There are significant negative correlations between PC1 and equatorial central Pacific (Fig. 2a, b). This relationship has been well documented in previous studies. The anomalous convection associated with central Pacific warming would produce a poleward extension of anomalous westerlies over WNP, leading to a strengthened and eastward displacement of monsoon trough (MT), which is in favor of TC genesis over WNP (Wang and Chan. 2002; Patricola et al. 2018; Wu et al. 2018; Zhang et al. 2020). The prevailing tracks of WNP TC also strongly modulated by the western Pacific subtropical high (WPSH) (Wang et al. 2013b; Camp et al. 2020) (Fig. 2c, d). The numbers of landfalling TCs would be reduced (enhanced) by intensified (weakend) WPSH. The prevailing TC tracks would be westward while WPSH is strong. When WPSH is weakened and shifted eastward, TC tracks can recurve northwards and affect East China, Korea and Japan. Moreover, positive VWS anomalies and over subtropical WNP while negative in mid-latitude is in favor of more TC activities (Fig. 2e, f). Negative VWS anomalies over WNP can be favorable atmospheric conditions for TC formation since more instabilities could be frequently developed in an organized convection system (Gray 1998).
To ensure the stability and robustness of the predictors, only predictors well above the 90% confidence level in both observational and model hindcast result was retained. For example, the SST of western Pacific warm pool were not selected because of the model’s poor ability in representing this relationship, although warm SST was in favor of TC genesis and intensification. The correlation between model hindcast and PC1 was comparatively weak on the whole. However, the spatial distribution of anomaly correlation coefficient (ACC) between PC1 and environmental fields are well captured by BCC_CSM1.1, despite discrepancies still exists, such as H850 in the SCS and East of Philippines (Fig. 2b). This discrepancy might be associated with weaken and northward shifted WPSH in model simulation, which is a common bias in CMIP5 models (He et al. 2014; Zhang et al. 2020). Three predictors were selected: SST anomalies in the central Pacific (10°S-10°N, 170°E-130°W), H850 over western and central north Pacific (10°N-30°N, 120°E-160°W), VWS over tropical WNP (0–15°N, 120°E-170°W).
The interannual variation of PC2 was connected to the canonical eastern Pacific ENSO (Fig. 2a) (Mei et al. 2015). The inverse variation of TC track density between SCS and WNP was associated with the eastward or westward shift of TC genesis location. Suppressed TC formation over northwest quadrant while enhanced TC generation over southeastern quadrant of WNP was associated both oceanic and atmospheric responses of El Nino forcing (Wang and Chan. 2002). Previous studies indicate that compared with Nino-3.4 indices, mega-ENSO index (Wang et al. 2013c) (Fig. 3a, b) exhibits higher predictive skill in seasonal forecasting of WNP TC activity (Zhan et al. 2017; Sun et al. 2020). Besides SST anomalies, the OLR of equatorial eastern Pacific (15°S-15°N, 140°W-90°W) associated with Walker circulation variation was selected as predictors.
The WNP tracks also exhibits significant inter-decadal variations with decreased westward and northeastward TC recurving tracks since early 2000s (Fig. 1c, 1d). Studies shows the modulation effect of Pacific decadal variability on both WNP TC genesis and prevailing tracks (Liu and Chan 2008; Zhao and Wu 2014; He et al. 2015; Liu et al. 2019). According to existed research and correlation analysis (Fig. 3a, 3b), Pacific decadal oscillation (PDO) index was taken into account to represent the influence of Pacific decadal variability on WNP TC activity, which was defined as the leading principal component of North Pacific (poleward of 20°N) monthly SST variability (Mantua et al. 1997).
The EOF3 of TC track density was associated with a north-south seesaw spatial pattern (Fig. 1e, f). Northward migration of mean latitudinal location of WNP TC genesis and tracks was revealed in recent studies (He et al. 2015; Wang and Toumi 2021). The ACC of eastern tropical Indian Ocean (20°S-10°N, 80°E-120°E) pass significant test at 90% confidence level and selected as predictor of PC3. Previous studies reveal that eastern Indian Ocean SST anomalies significantly affect both summer monsoon and equatorial Kelvin wave activity over western Pacific and modulate TC activity (Zhan et al. 2011; Ha et al. 2015). Besides, the correlation between SST anomalies of Kuroshio and its extension (K-KE) area (25°S-45°N, 140°E-150°W) was significant and thus selected as predictors(Fig. 4a, 4b). The positive SST anomaly in K-KE was favorable for the warm high over Northeast Asia and Japan, leading to the northward displacement of WPSH (Ding et al. 2019), give rise to more northward TC over WNP (Camp et al. 2019, 2020).
4.2 Predictive skill of BCC_CSM1.1 for selected background environmental fields
The predictive skill of BCC_CSM1.1 for large scale environmental fields associated with WNP TC was evaluated, including SST, H850, VWS and OLR (Fig. 5). The distribution of anomaly correlations between the observed and model hindcasts for JAS over 1991–2020 initated in June (1-month lead) was given. The prediction skill of interannual variation of SST in JAS is the highest, with ACC exceeds 0.8 in equatorial Pacific (Fig. 5a). BCC_CSM1.1 also well captures the interannual variability of H850, with ACC for most regions between 30°N and 30°S passed significant test at 95% confidence level. For eastern Indian Ocean, southern warm pool and parts of east equatorial Pacific, the ACC reaches 0.8 (Fig. 5b). Significant skill for VWS is shown over the warm pool and central equatorial Pacific, where the VWS is an important factor for WNP TC activity. Reduced VWS and eastward extension of monsoon trough provides a favorable condition for TC genesis (Fig. 5c) (Wu et al. 2012; Cao et al. 2016). Significant skill for OLR can be found over equatorial central and eastern Pacific, subtropical northeastern Pacific and parts of western Pacific warm pool, with ACC reaching 0.8 in equatorial eastern Pacific (Fig. 5d).
4.3 Validation for WNP TC track density prediction
The procedure of forecast model construction can be divided into three steps. The first step is multi-linear regression between time series of PCs and selected predictors for each leading mode of EOF analysis. Then the regression coefficient of training period was applied to the forecast of target year to obtain forecast PCs. Finally, the spatial distribution of track density was reconstructed by composite of leading EOF modes forecasted multiplied by variance explained.
Skillful forecast of PCs is vital for the overall performance of TC track density prediction. The regression analysis is first applied to the whole hindcast period to check the ability of selected predictors to reconstruct leading PCs. Figure 6 shows the regressed time series of PCs based on selected predictors of both observational and model hindcast results. The interannual variation of three leading PC are well captured by regressed models. ACC between observational and regressed PCs passed significant test at 99% confidence level, with PC1 based on observational predictors ranks the highest.
On the other hand, to test the prediction skill of forecast model in real time forecast, one-year-out cross validation was utilized (Fig. 7). For random chosen target year of prediction among 1991–2020, the other 29 years was utilized as training set for both observational and model hindcast. Regression analysis and forecast model construction based on training set was repeated for 30 times until all years have been chosen as target years of prediction. As shown in Fig. 7, the interannual variation of reforecast PCs is in good agreement with observations. The ACC between observational and reforecast based BCC_CSM1.1 for three PCs are 0.52, 0.54 and 0.62, all passed significant test at 99% confidence level.
TC activity over WNP and SCS is regional dependent. To examine the spatial distribution of prediction skill of TC track density, WNP (0–45°N, 120°E-180) is partitioned into 4 sub-regions with 20°N and 140°E serve as the dividing line between the south and north and between the east and west following other studies (Wang and Chan 2002; Wang et al. 2013a) but with slight modification (Fig. 8). Figure 9 shows that forecast for SCS and southeast quadrant of WNP exhibits higher skill, with ACC of forecast initiated in June exceeds 0.6.
The relationship between prediction skill and initial date of hybrid model was evaluated. The prediction skill shows an overall increasing trend with decreased forecast lead time, indicates the source of predictability of TC tracks originates not only external forcing, but also initial state of atmospheric-ocean coupled system. The spatial distribution of ACC for different initial date was given (Fig. 10). The interannual variation TC activity in southeast quadrant of WNP was captured well. It is also noted that in the forecast initiated in April and May, TC tracks passing East China Sea and SCS are in good agreement with those from observations.
As the strongest signal in air-sea system, ENSO is an important source of predictability in seasonal to inter-annual timescale (Wang et al 2013a, b). Wang and Chan (2002) noted that WNP TC activities in July-December are noticeably predictable under the background of strong ENSO events in preceding winter and spring. The relationship between predictive skill of WNP TC track density and different phases of ENSO was investigated. First, all El Niño and La Niña events from 1991 and 2020 was determined based on Niño 3.4 index (Ren et al. 2018). El Niño years are 1991/92, 1994/95, 1997/98, 2002/03, 2004/05, 2006/07, 2009/10, 2014/15, 2015/16, 2018/19. La Niña years are 1995/96, 1998/99, 1999/00, 2000/01, 2005/06, 2007/08, 2008/09, 2010/11, 2011/12, 2016/17, 2017/18, 2020/2021. Due to the phase lock phenomenon of ENSO events, i.e. most El Niño (La Niña) events peak in boreal winter, all years was classified into El Niño (La Niña) developing, decaying or neutral phases. In order to ensure the reliability of the assessment, all 24 ensemble members of model hindcasts were utilized. The averaged ACC of all forecast members was used to represent the prediction skill of hybrid model.
The prediction skill of hybrid model was investigated according to different ENSO phases. Forecast skills for the El Niño decaying years is higher than those for both neutral and La Niña decaying years, especially in southeast quadrant of WNP (Fig. 11). The southwest quadrant of WNP shows higher skill in neutral years, while the southeast quadrant is the highest in La Niña decaying years. The forecast of TC track density in ENSO developing years is generally less skillful compared with decaying years (Fig. 12). Relative to El Niño, the La Niña developing years show a slightly higher skill. This might be associated with the fact that most La Niña developed in El Niño years. Meanwhile, the regions with ACC passing significant test was mainly located in SCS. Since the relationship between ENSO and WNP TC activity depends on the intensity of ENSO events, only strong El Niño (La Niña) events have significant influence on WNP TCs (Wang and Chan 2002). The intensity of ENSO events shows a remarkable positive asymmetry between its two phases, i.e. the strongest El Niño is stronger than the strongest La Niña (An and Jin 2004; Liang et al. 2017). This disparity in prediction skill indicates the source of predictability of WNP TC tracks might be originate from ENSO events, especially strong El Niño events.